Post

Optimizing blur hash, step by step: www.romainguy.dev/posts/2024/o...

romainguy.devOptimization, Step by StepBlurHash is a compact representation of placeholders for images. A blur hash is encoded as a short string that can be rendered to a bitmap at runtime to display a “blurry” version of the source image....

26/11/24, 7:06 am26 reposts111 likes

Mārtiņš Možeiko

@martins.ninja

26 Nov 2024

How about replacing sRGB/linear conversions with simple square & sqrt(x)? I'd guess that for these kind of images the difference will be very minimal, but the code will be smaller (no tables) and probably as fast. And easier to apply SIMD.

Romain Guy

@romainguy.dev

26 Nov 2024

I gave it a try and on mobile it makes things worse, losing 20% of performance. sqrt() is unfortunately still fairly expensive.

Romain Guy

@romainguy.dev

26 Nov 2024

That's also a possibility yes. If I could write SIMD easily in Kotlin it's definitely the direction I'd take.

Sebastian Aigner 📡

@sebi.io

26 Nov 2024

I don't have experience with this but I'm very curious since you also mentioned it in your blog post; is it possible to write Kotlin code that makes use of SIMD at all? Or does that necessary require to drop down into native code via NDK?

Romain Guy

@romainguy.dev

26 Nov 2024

Some vectorization can happen inside ART, I haven't checked what it would take to ensure this particular algorithm benefits.

Roy van Rijn

@royvanrijn.com

26 Nov 2024

One option would be to use Kotlin-native, tweak the code and check the result. In Java you can soon use vectorization (JEP-448) and do everything more easily, especially powerful if you combine it with Graal native compilation and their inferred optimizations.